Replicate MCP server for your AI

Create an empty model on this account, usually as a training destination.

List the models in one curated collection.

Get one model: description, run_count, whether it is official, its latest version id, and the inputs it takes as a flat list.

See what a model takes before running it, because every model has different inputs.

Read a model's README as Markdown, for usage advice, limits and licence notes the input list does not explain.

List Replicate's curated model collections, like super-resolution or image-to-text.

List example runs the author saved for a model.

List a model's versions, newest first, with id and date.

Browse all public models, newest version first by default.

The best first step for finding a model: search public models, collections and docs by what you want to do, like "remove image background".

Plain text search over public models, page by page, when you want more results than replicate_search_models gives.

Change a model's description, readme, or its github, paper, weights or license links.

Stop a run that is still going.

Start a run by version.

Check one run once: status (starting, processing, succeeded, failed, canceled, aborted), output, output_urls, error, timings and the last lines of its logs.

List this account's runs, newest first, 100 per page, with status, model, times and output.

Run a model through one of this account's deployments and get the result.

Run any AI model and get the result: the normal way to generate an image, video, audio or text.

Wait for a run that is still going, and get its result when it ends.

Create a deployment: a private endpoint for one model version.

Get one deployment with its current release: model, version, hardware and instance counts.

List this account's deployments (private endpoints with chosen hardware and scaling), newest first, each with its current model, version, hardware and instance counts.

List the hardware Replicate offers, as name and sku.

Change a deployment's hardware, scaling or model version.

Stop a training that is still going.

Delete an uploaded file.

Read a file from a signed Replicate download link, for a link that holds owner, expiry and signature.

Get one uploaded file: size, type, checksum, metadata, expiry and urls.get, the link to pass as a model input.

Check one training: status, error, timings, the last lines of logs, and output.version and output.weights when done.

List files uploaded to this account through the API, newest first, with id, size, type, expiry and the link to use as a model input.

List this account's trainings, newest first, 100 per page, with status, model, times and the new version when done.

Start training a new version of a model on your own data, usually a fine-tune.

Upload a file to Replicate so a model can read it, when it has no public link or is too big for a data URL.

Get the Replicate account this connection uses: type (user or organization), username, name.

Delete a model.

Delete a model version and all its runs and output files.

Get one model version with its full OpenAPI schema.

Start a run on one of this account's deployments.

Start a run of an official model, by owner and name with no version id.

Delete a deployment.

Start training from a specific model version.

Read the signing secret that proves a webhook really came from Replicate.